Improve Hugging Face integration

[NielsRogge]

Hi @zhaorw02, thanks for making the model available on 🤗 !

This PR improves the Hugging Face integration by:

• equiping the model with from_pretrained and push_to_hub capabilities
• use safetensors for weights serialization
• have download stats for your model.

This is all supported thanks to the PyTorchModelHubMixin class. It works like so:

from lit_gpt.model_cache import GPTCache

model = GPTCache(...)

# equip with weights
filepath = hf_hub_download(repo_id="zzzrw/DeepMesh", filename="pytorch_model.bin")
state_dict = torch.load(filepath, map_location="cpu")
model.load_state_dict(state_dict)

# push to the hub
model.push_to_hub("zzzrw/DeepMesh")

# now anyone can use it like so:
model = GPTCache.from_pretrained("zzzrw/DeepMesh")

Feel free to try it out!